From Panthor to RK3588: Advancing graphics, video and SoC support in Linux kernel 7.0

Loic Molinari
April 13, 2026

The Linux kernel 7.0 has been released! While the major version bump reflects a progression driven by Linus Torvalds' versioning preferences rather than a major overhaul, this release includes several noteworthy additions in hardware support, performance, and security.

Here are some of the highlights we are excited about:

• Enablement work for upcoming hardware including the Intel Nova Lake platforms, the next-generation Intel Xeon Diamond Rapids and AMD Zen 6 processors.
• Improved performance for computer graphics workloads on Mali and NVIDIA GPUs thanks to larger pages support (see details below for Mali).
• Rust programming language is now a core part of the kernel with the experimental tag being removed thanks to great work done by the Rust for Linux team.

Once again, we recommend the LWN's articles to get an overview of all the changes in this development cycle: part 1 and part 2.

Collabora contributed to Kernel 7.0 with no less than 227 patches authored by 24 developers. Let's have a look at all of these contributions!

Graphics

Boris Brezillon added support for atomic page table updates to the Panthor driver, which will be needed for Vulkan sparse binding support. He also fixed a long standing scheduler bug causing GPU contexts to not be scheduled properly when the number of GPU contexts exceeds the number of firmware slots, along with a bunch of other fixes. He worked conjointly with Faith Ekstrand and Loïc Molinari on exposing support for CPU-cached mappings on IO-incoherent systems to speed up accesses of GPU buffers from the CPU.

Loïc also proposed patches to reduce the page table overhead of DRM drivers on systems using transparent huge pages. The shared memory backing store will now attempt 2 MiB-aligned and contiguous allocations whenever possible. This speeds up both CPU and GPU mappings of memory. This is enabled by default for the Panfrost/Panthor, V3D, and i915 DRM drivers. For Panthor on a Rock 5B, this patch series makes the first memcpy to an entire BO mapped in userspace about twice as fast.

Continuing with huge pages improvements, Adrián Larumbe proposed a patch that avoids splitting huge pages in order to improve performance for frequent partial evictions.

Ashley Smith added timestamp propagation to the Panthor driver so that more accurate timestamps can be retrieved via perf events or via the SYNCOBJ_WAIT ioctl.

Gert Wollny added support for running a PPU flop reset on the GPU GC8000 Nano Ultra VIP r6205 (etnaviv DRM driver) as found on systems such as the ST STM32MP25. It is required to get correct rendering results after powering up from resume.

Deborah Brouwer proposed code cleanups, warning fixes, and modernization in Tyr, the Rust DRM driver for Mali GPUs.

Multimedia

Benjamin Gaignard continued to improve the Verisilicon AV1 stateless decoder stability and conformance by fixing bugs reported by our customers for specific and rare AV1 stream configurations.

NXP have started working on moving to upstream stateless decoder drivers for their BSP and have been facing hardware issues on IM8MQ, where running the two VPU cores concurrently causes memory corruption. Nicolas Dufresne has been helping by turning the v4l2_m2m_dev (which is in fact a scheduler) into a ref-counted object, so it can be shared by the two VPU devices. This allows doing event-based time-multiplexing and prevents both cores from running concurrently.

Dmitry Osipenko proposed a fix to detect and report non-working HDMI input interrupt for the Synopsis HDMI Receiver when using Rockchip's binary-only trusted firmware instead of the open source version on the Rockchip RK3588.

Rockchip hardware enablement

Cristian Ciocaltea expanded the common HDMI PHY configuration API to include the required options for HDMI 2.1 FRL and wired this up in the Rockchip Samsung HDPTX PHY driver, used on RK3588/RK3576 SoCs. HDMI 2.1 transitions from legacy TMDS to Fixed Rate Link (FRL) signaling and achieves a massive 48 Gbps of bandwidth, unlocking lower latency and ultra-high resolutions like 8K@60Hz and 4K@120Hz. Note that support for FRL on the Rockchip SoCs also requires changes in the HDMI controller driver, which have not yet landed. He also landed a work-around for a limitation of the PHY PLL calculator responsible for breaking the video output when using a particular 1080p@120Hz display mode with 10 bits-per-channel and he improved the hotplug detection reliability in the Rockchip DW HDMI QP driver.

Detlev Casanova added support for the H264 and HEVC decoders on RK3588 and RK3576 SoCs. It includes adding an extension to the V4L2 stateless decoder API to support long and short term reference frame sets data. Some rework of the driver was done to support less similar variants than before and extract common functions for all variants. Multicore support is not there yet, but some preparations were done to probe each core in the same instance of the driver and use struct-based register maps to allow preparing decode runs before knowing which core will be used.

Sebastian Reichel prepared the S/PDIF driver used by Rockchip RK3588/RK3576 SoCs for DisplayPort audio support. In the meantime, he also proposed a couple of cleanup patches for that driver. Support for DisplayPort audio itself is being worked on, but has not yet landed.

Michael Riesch continued improving video capture and camera support for Rockchip RK35 SoCs by adding a driver for the MIPI CSI-2 receiver unit. Check out the recording of his FOSDEM '26 presentation for a comprehensive overview.

Daniel Stone fixed up the conformance of the VOP2 display controller with KMS. Atomic checks can now properly fail when applying an unsupported state instead of silently working around hardware limitations.

MediaTek Hardware Enablement

AngeloGioacchino Del Regno expanded support for several MediaTek SoCs like the MT8196 Chromebook including power management enablements (SPMI, DVFSRC, and regulator), display and audio enablements (HDMI, DSI and audio controller), DMA fixes, and Device Tree cleanups.

Louis-Alexis Eyraud enabled the HDMI output for the Mediatek Genio 1200-EVK board and HMDI sound output support for the Mediatek Genio 510, 700, 1200-EVK, and Radxa NIO-12L boards. He also added a new entry in the MediaTek Soc Info driver data for the MT8371 SoC (Genio 520). Finally he added new compatibles in the Device Tree bindings for SPMI and I2C on MT8189 SoC and for HDMI PHY IP on MT8188 SoC.

Nicolas Frattaroli contributed preliminary patches to enable support for the MT8196 SoC. He fixed clock issues and read enabled cores from eFuse to activate the GPU (this is good progress but the GPU will stay off in mainline for stability reasons for now). He also started refactoring and fixing the interconnect driver.

Continuing enablement for the MT8196 SoC, Laura Nao introduced support for the LVTS thermal controller including the necessary driver changes for handling specific conversion logic and calibration offsets.

Nicolas Dufresne improved the Mediatek stateless VP9 decoder to reject upfront decode operation where the subsampling is unsupported by the hardware (YUV 422 and 444).

Sebastian Fricke contributed a patch for the MediaTek Video Codec, preparing for proper error propagation and multi-frame handling.

Nícolas F. R. A. Prado added extra clocks documentation for the MT8195.

Miscellaneous bits

And as usual, all of the above is accompanied by a mixed bag of fixes and improvements.

Beyond his Rockchip work, Cristian Ciocaltea contributed several stability fixes and cleanups for the NAU8821 audio codec driver, which powers devices like the Valve Steam Deck.

Derek Foreman extended the perf tool with a new option to convert events within a specified time range.

Sjoerd Simons proposed a first set of patches upstreaming support for the OpenWrt One, an open hardware access point.

And finally, Vivek Das Mohapatra landed an AMDGPU DRM fix to initialize the backlight level values from the hardware.

Below is a full list of contributions made by Collabora for the 7.0 release, as recorded in the git commit history: